ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 20:09:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

GEMM+Bias+ReLU+Add (#76)

Browse Source 
* tweak conv for odd C

* update script

* clean up elementwise op

* fix build

* clean up

* added example for gemm+bias+relu+add

* added example for gemm+bias+relu

* add profiler for gemm_s_shuffle; re-org files

* add profiler

* fix build

* clean up

* clean up

* clean up

* fix build

[ROCm/composable_kernel commit: 823657ed12]
This commit is contained in:
Chao Liu
2022-02-06 22:32:47 -06:00
committed by GitHub
parent 8890cc207d
commit 8efcb80fa5
77 changed files with 3865 additions and 932 deletions
Download Patch File Download Diff File Expand all files Collapse all files

101
profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp
View File

@@ -3,11 +3,11 @@
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_fwd_bias_activation_add.hpp"
#include "element_wise_operation.hpp"
#include "device_conv_fwd_bias_activation_add.hpp"
#include "reference_conv_fwd_bias_activation_add.hpp"
namespace ck {
namespace tensor_operation {
@@ -30,56 +30,6 @@ void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instan
namespace ck {
namespace profiler {
template <typename TIn,
typename TWei,
typename TOut,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp>
void host_reference_calculation(const Tensor<TIn>& in_n_c_hi_wi,
const Tensor<TWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_k_ho_wo,
const Tensor<TOut>& bias_k,
const Tensor<TOut>& resi_n_k_ho_wo,
const std::vector<ck::index_t>& conv_strides,
const std::vector<ck::index_t>& conv_dilations,
const std::vector<ck::index_t>& in_left_pads,
const std::vector<ck::index_t>& /* in_right_pads */,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
for(int c = 0; c < wei_k_c_y_x.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei_k_c_y_x.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * conv_strides[0] + y * conv_dilations[0] - in_left_pads[0];
for(int x = 0; x < wei_k_c_y_x.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * conv_strides[1] + x * conv_dilations[1] - in_left_pads[1];
if(hi >= 0 && hi < in_n_c_hi_wi.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in_n_c_hi_wi.mDesc.GetLengths()[3])
{
v += in_element_op(static_cast<const double>(in_n_c_hi_wi(n, c, hi, wi))) *
wei_element_op(static_cast<const double>(wei_k_c_y_x(k, c, y, x)));
}
}
}
}
out_n_k_ho_wo(n, k, ho, wo) = out_element_op(v, bias_k(k), resi_n_k_ho_wo(n, k, ho, wo));
};
make_ParallelTensorFunctor(f_nchw,
out_n_k_ho_wo.mDesc.GetLengths()[0],
out_n_k_ho_wo.mDesc.GetLengths()[1],
out_n_k_ho_wo.mDesc.GetLengths()[2],
out_n_k_ho_wo.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
@@ -169,20 +119,37 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
host_reference_calculation(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
resi_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
using ReferenceConvFwdInstance =
ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation_Add<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
resi_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
@@ -240,9 +207,9 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();

131
profiler/include/profile_conv_fwd_bias_relu_impl.hpp
View File

@@ -3,11 +3,11 @@
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_fwd_bias_activation.hpp"
#include "element_wise_operation.hpp"
#include "device_conv_fwd_bias_activation.hpp"
#include "reference_conv_fwd_bias_activation.hpp"
namespace ck {
namespace tensor_operation {
@@ -30,84 +30,6 @@ void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances(
namespace ck {
namespace profiler {
void cpu_conv_bias_relu(ck::half_t* in_ptr,
ck::half_t* weight_ptr,
ck::half_t* output_ptr,
ck::half_t* bias_ptr,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const ck::index_t Y,
const ck::index_t X,
const ck::index_t Hi,
const ck::index_t Wi,
const ck::index_t Ho,
const ck::index_t Wo,
const ck::index_t Stride,
const ck::index_t Dilation,
const ck::index_t Pad)
{
const auto in_desc =
HostTensorDescriptor(std::vector<std::size_t>{static_cast<std::size_t>(N),
static_cast<std::size_t>(Hi),
static_cast<std::size_t>(Wi),
static_cast<std::size_t>(C)});
const auto wei_desc =
HostTensorDescriptor(std::vector<std::size_t>{static_cast<std::size_t>(K),
static_cast<std::size_t>(Y),
static_cast<std::size_t>(X),
static_cast<std::size_t>(C)});
const auto out_desc =
HostTensorDescriptor(std::vector<std::size_t>{static_cast<std::size_t>(N),
static_cast<std::size_t>(Ho),
static_cast<std::size_t>(Wo),
static_cast<std::size_t>(K)});
const auto bias_desc =
HostTensorDescriptor(std::vector<std::size_t>{static_cast<std::size_t>(K)});
auto f_k = [&](auto k) {
for(int n = 0; n < N; ++n)
{
for(int ho = 0; ho < Ho; ++ho)
{
for(int wo = 0; wo < Wo; ++wo)
{
double v = 0;
for(int c = 0; c < C; ++c)
{
for(int y = 0; y < Y; ++y)
{
int hi = ho * Stride + y * Dilation - Pad;
for(int x = 0; x < X; ++x)
{
int wi = wo * Stride + x * Dilation - Pad;
if(hi >= 0 && hi < Hi && wi >= 0 && wi < Wi)
{
double in =
in_ptr[in_desc.GetOffsetFromMultiIndex(n, hi, wi, c)];
double wei =
weight_ptr[wei_desc.GetOffsetFromMultiIndex(k, y, x, c)];
v += in * wei;
}
}
}
}
v += bias_ptr[bias_desc.GetOffsetFromMultiIndex(k)];
v = v > 0 ? v : 0;
output_ptr[out_desc.GetOffsetFromMultiIndex(n, ho, wo, k)] = v;
}
}
}
};
make_ParallelTensorFunctor(f_k, K)(std::thread::hardware_concurrency());
}
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
@@ -191,24 +113,35 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
cpu_conv_bias_relu(in_n_c_hi_wi.mData.data(),
wei_k_c_y_x.mData.data(),
out_n_k_ho_wo_host_result.mData.data(),
bias_k.mData.data(),
N,
K,
C,
Y,
X,
Hi,
Wi,
Ho,
Wo,
conv_filter_strides[0],
conv_filter_dilations[0],
input_left_pads[0]);
using ReferenceConvFwdInstance =
ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
@@ -263,9 +196,9 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();

44
profiler/include/profile_conv_fwd_impl.hpp
View File

@@ -3,11 +3,11 @@
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_fwd.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_fwd.hpp"
namespace ck {
namespace tensor_operation {
@@ -105,15 +105,37 @@ void profile_conv_fwd_impl(int do_verification,
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
host_conv_nchw_kcyx_nkhw(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
using ReferenceConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvFwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
@@ -177,9 +199,9 @@ void profile_conv_fwd_impl(int do_verification,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();

286
profiler/include/profile_gemm_bias_relu_add_impl.hpp Normal file
View File

@@ -0,0 +1,286 @@
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm_bias_activation_add.hpp"
#include "reference_gemm_bias_activation_add.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmBiasReluAddPtr = ck::tensor_operation::device::DeviceGemmBiasActivationAddPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddReluAdd>;
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_relu_add_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int StrideC1,
int KBatch = 1)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
// c1_m_n[m ,n]
Tensor<BDataType> c1_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::cout << "c1_m_n: " << c1_m_n.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
c1_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
c1_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{});
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddReluAdd;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivationAdd<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
b_k_n,
c_m_n_host_result,
c0_n,
c1_m_n,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
DeviceMem c1_m_n_device_buf(sizeof(CDataType) * c1_m_n.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
c1_m_n_device_buf.ToDevice(c1_m_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasReluAddPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_nk_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instances(
gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c1_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op,
KBatch);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * N +
sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
check_error(c_m_n_host_result, c_m_n_device_result);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a: ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c0: ", c0_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c1: ", c1_m_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck

264
profiler/include/profile_gemm_bias_relu_impl.hpp Normal file
View File

@@ -0,0 +1,264 @@
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm_bias_activation.hpp"
#include "reference_gemm_bias_activation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmBiasReluPtr = ck::tensor_operation::device::DeviceGemmBiasActivationPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddRelu>;
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_relu_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int KBatch = 1)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddRelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, c0_n, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasReluPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
KBatch);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
check_error(c_m_n_host_result, c_m_n_device_result);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c0 : ", c0_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck

55
profiler/include/profile_gemm_impl.hpp
View File

@@ -1,4 +1,14 @@
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm.hpp"
#include "reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
@@ -15,6 +25,11 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNo
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
@@ -86,17 +101,30 @@ void profile_gemm_impl(int do_verification,
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
host_gemm_mk_kn_mn(a_m_k,
b_k_n,
c_m_n_host_result,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
@@ -184,6 +212,9 @@ void profile_gemm_impl(int do_verification,
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
@@ -191,6 +222,9 @@ void profile_gemm_impl(int do_verification,
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
@@ -198,6 +232,9 @@ void profile_gemm_impl(int do_verification,
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
@@ -205,6 +242,9 @@ void profile_gemm_impl(int do_verification,
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
@@ -283,8 +323,7 @@ void profile_gemm_impl(int do_verification,
}
else
{
std::cout << "this device GEMM instance does not support this GEMM problem"
<< std::endl;
std::cout << "does not support this GEMM problem" << std::endl;
}
}